Bruce Cumming: How the disparity selective neurons adapt to the statistics of the environment
Talk by Bruce Cumming of the National Institutes of Health. Given to the Redwood Center for Theoretical Neuroscience at UC Berkeley.
To generate an efficient representation, cortical processing should exploit knowledge of the statistical properties of the outside world. Binocular vision is a useful system to explore this specialization, because the projection geometry imposes some very strong statistical regularities. The responses of disparity selective neurons in the primate striate cortex reflect some of these constraints. When one eye is presented with an image that is the photographic negative of the other eye’s image (anticorrelation), these neurons still respond to disparity, but these responses are attenuated relative to those elicited by correlated stereograms. This attenuation represents an adaptation to the statistics of the outside world (anticorrelation does not occur in natural viewing). Here we investigate the mechanisms that generates this attenuation. One simple explanation, compatible with all existing data, is that these neurons reflect the summation of multiple subunits, each of which behaves like the binocular energy model. To test this hypothesis, we employed spike-triggered analysis of covariance, a technique designed to reveal the properties of such summed elements. This revealed that most disparity selective neurons are indeed composed of multiple subunits, but that their spatial organization did not explain the observed attenuation in response to anticorrelated stimuli.
The spatial properties suggested that attenuation might be explained if the subunits, rather than summing linearly, were involved in a recurrent processing loop. If recurrent processing is required, then the earliest part of the response to anticorrelated stimuli should not show attenuation. We examined the dynamics of the disparity response with subspace mapping. Random dot patterns were presented in which the disparity changed randomly every 30ms. Whether each disparity was correlated or anticorrelated was also randomized. We then examined the timecourse of mean responses to each disparity. In the population mean, the response to disparity started with a latency of 40ms. For the next ~15ms responses to correlated and anticorrelated disparities were of nearly equal magnitude. After this point, the extent of attenuation increased steadily over a further 15ms, after which is remained approximately constant.
This delayed attenuation might in principle be explained by a simple nonlinearity. For example, a rectifying nonlinearity would only have its effect once the rates (for some stimuli) reached zero. If the attenuation is explained by a nonlinear function of rate, then during the falling phase of the response, when the response rates fall to the values observed from 40-65ms, the attenuation should also decrease. In fact the attenuation remained constant throughout the falling phase of the response, excluding explanations based on simple output nonlinearities.
Thus the attenuation of responses to anticorrelation develops more slowly than the response to disparity, in a way that is not explained by simple nonlinearity. This suggests that recurrent processing plays a role in generating this adaptation to the statistics of binocular vision.